Development and validation of a deep learning-based model for estimating stroke volume and blood pressures...

Cardiac output (CO) is essential for systemic oxygen delivery, but gold-standard measurement using pulmonary artery catheter thermodilution remains invasive. Arterial pressure-based cardiac output (APCO) devices are widely used but still require arterial catheterization. We developed and validated a deep learning (DL) model that simultaneously estimates stroke volume (SV), systolic blood pressure (SBP), and diastolic blood pressure (DBP) using electrocardiogram (ECG) and photoplethysmogram (PPG).

We retrospectively analyzed 881 surgical cases at Seoul National University Hospital (Aug 2016 - Dec 2020) with simultaneous ECG, PPG, arterial blood pressure waveform (ABP), SV measured by the APCO devices, and demographics. After signal-quality rules and arterial-line plausibility filters, 20-s ECG/PPG segments were paired SV to form samples. A DL model using ECG, PPG, the first and second derivatives of PPG, demographic variables, and ECG-derived features simultaneously estimated SV, SBP, and DBP. Model performance was evaluated in a temporal hold-out cohort using error metrics, Bland-Altman analysis, and 10-min interval trend analysis. SBP and DBP performance was assessed against ISO 81060-3:2022 criteria.

The test set comprised 128 cases (31,637 segments). Our DL model achieved a mean absolute error of 10.761 mL/beat and percentage limits of agreement ranging from -26.905% to 29.563%. SBP and DBP showed mean errors of 0.661 and 1.439 mmHg, with standard deviations of 9.716 and 6.684 mmHg, respectively, meeting the ISO 81060-3:2022 criteria.

A DL model using ECG/PPG and demographics can estimate SV with clinically useful performance while simultaneously predicting SBP/DBP, potentially expanding continuous hemodynamic monitoring to non-arterial-line cases.